Fake News as Discursive Integration: An Analysis Of Sites That Publish False, Misleading, Hyperpartisan And Sensational Information
Description
Codes for content analysis of stories from sites that were labelled fake news during the 2016 campaign. Codebook, syntax and stories can be provided upon request. Data Analysis RQ 1 (How do fake news sites employ elements of a) misinformation, b) bias, c) clickbait and d) sensationalism?) was answered using descriptive statistics and correlations measuring the degree to which each of these concepts related to each other in the content analyzed. Two sets of regression models were used to assess the impact of sensationalism (RQ2), clickbait (RQ3), misinformation (H1) and bias (H2) on engagement outcomes on Facebook and Twitter. In this dataset, more than one story published by the same outlet was selected for coding, violating the assumption of independent observations. Thus, our data are nested and we might expect that two stories from the same outlet will tend to be more alike than two stories selected from different outlets. This violation was controlled for by using linear mixed-effects model, also referred to as two-level hierarchical or multilevel model (Steele, 2008). Because of the skewed nature of the dependent variable (social media engagement levels), the distribution of residuals was not normal, and log-transformations were performed on those variables following the recommendation of Tabachnick and Fidell (2007). In order to control for the nested character of this data, two sets of mixed-effects models were conducted using MIXED on SPSS and MuMin in R. The independent variables included: Trump valence, Clinton valence, clickbait, sensationalism, misinformation and strength of partisanship (folded bias variable). The original bias variable was removed from this regression analysis because it was collinear with both candidates’ valence and its folded counterpart. The model controls for “random effects,” that is, the mere fact of a story belonging to a specific outlet. After inspecting the plots for the observed relationships, we identified a curvilinear relationship between strength of partisanship and engagement outcomes. A new block including the squared variable for strength of partisanship was entered in the analysis, which improved significantly the model fit for the analyses based on Akaiake’s Information Criterion. In the case of mixed-effects models, two types of R2 can be calculated to describe variance explained: marginal and conditional. The former describes the portion explained by fixed effects (content characteristics) and the latter shows the variance explained by both fixed (content characteristics) and random (published by a specific outlet) effects (Nakagawa and Schielzeth 2012).